We shipped Quobyte 5 last month, and this week we get to show it off in the city where Quobyte started! AI workloads shift quickly, from training to inference, and Quobyte stays fast across all of them. It can run directly on your GPU nodes, turning their spare CPU and NVMe into high-performance storage. Quobyte’s bulletproof resiliency handles drive, node, and rack failures in software, keeping your file system available with 100% uptime. That’s Storage Architected for AI™. Through Thursday, the Quobyte team is in the Presidential Suite at The Posthouse Berlin Potsdamer Platz, five minutes away from GTC Berlin, ready to walk you through the architecture and all that is new in Quobyte 5. Book a meeting: quobyte.com/gtcberlin/ #GTCBerlin #AIInfrastructure #AIStorage #hpc

ICYMI: the Inside Quobyte 5 webinar is now on demand. Updating to Quobyte 5 got 5x the metadata operations out of the same previous-generation servers, around 500,000 per server, plus up to 10x faster I/O on IOR hard writes. Björn Kolbeck, Quobyte CEO and co-founder, explains why this release focuses on performance. Hardware lead times are long, prices keep climbing, and some quotes now expire in 24 hours, so getting more out of the servers already in your data center has become a necessity. Also covered: • CPU topology optimization that adapts to your servers with no configuration • Network link aggregation across TCP, RoCE, and InfiniBand, with no bonding required • GPU Converged Storage that recovers on its own when GPU nodes crash • Asking your file system questions in plain language through the new MCP server interface The Q&A answers the questions buyers ask most: does a large metadata query slow down production, how much CPU does Quobyte use on a GPU node, and does any metadata leave your cluster when you connect an LLM? Existing Quobyte 4 clusters get the boost through a standard rolling update with no downtime. Watch: buff.ly/vxH9yT0 #HPC #AIInfrastructure #SoftwareDefinedStorage

There is more storage performance in the servers you already run. Tomorrow, we're showing how Quobyte 5 puts it to work. By scaling across modern CPU cores and driving multiple InfiniBand or RoCE links in the same server, Quobyte 5 delivers substantially higher I/O and metadata performance. Your workloads run faster on the same servers, with no new hardware to buy and no added power draw. Only a few spots remain. Inside Quobyte 5: Turn Your Hardware Into Lightning-Fast AI Storage Tomorrow, October 7th, 10 AM PT Save your seat: buff.ly/fCRKv99 #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage #Quobyte

The usual answer to an AI workload outgrowing its storage is to buy more hardware. Then you find out the lead times, what another rack does to your power budget, and that none of it helps if storage was the bottleneck in the first place. Quobyte 5 is built to push what you already own further before you reach that point. It delivers substantially higher I/O and metadata performance on the servers and GPU nodes you already run. Modern server CPUs deliver most of their gains through core count, and Quobyte 5 scales across those cores on current and previous-generation systems alike. On the network side, RDMA multi-link with our link aggregation lets multiple InfiniBand or RoCE links in the same server work together, which matters most on dense GPU nodes with four or more high-speed interfaces. The practical result: the same workload runs on fewer servers already in your racks and uses less power. Read more: buff.ly/Jfj6moG #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage

The fastest storage upgrade you can make this quarter has no lead time and adds nothing to your power budget. It runs on the servers you already have. On October 7th, we're showing how Quobyte 5 does it. Quobyte 5 scales across the high core counts of modern CPUs and drives multiple InfiniBand or RoCE links in the same server, delivering substantially higher I/O and metadata performance from the x86 and ARM servers already in your racks. Your workloads run faster on the same servers, with no new hardware to buy and no added power draw We'll close with a live look at the new MCP server interface, where you can ask your file system questions in plain language. Inside Quobyte 5: Turn Your Hardware Into Lightning-Fast AI Storage October 7th, 10 AM PT Register: buff.ly/fCRKv99 #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage #Quobyte

Adding hardware to your infrastructure takes months, raises your power draw, and does nothing if storage is still the bottleneck keeping your GPUs waiting. On October 7th, we're walking through what Quobyte 5 does about that. We’ll cover how Quobyte 5 gets substantially higher I/O and metadata performance from the x86 and ARM servers you already run, how it uses modern CPUs and multiple InfiniBand or RoCE links, and how GPU nodes fit into the same architecture. We’ll also demo how you can ask your file system a question in plain language and work directly with live metadata. The question we’ll start with: Are we using our flash tier efficiently? How much data could we shift to HDD? You’ll see how the LLM works through the underlying metadata and how quickly you can drill down from there. Inside Quobyte 5: Turn Your Hardware Into Lightning-Fast AI Storage October 7th, 10 AM PT Register: buff.ly/fCRKv99 #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage #Quobyte

Every storage team has a version of this conversation: where did all our capacity go, and can we get some of it back? Finding the answer takes too much work, so the question often gets dropped. Quobyte 5's MCP server interface changes the cost of asking. You can start with a question in plain language, get answers directly from live file-system metadata in minutes, and drill down from there. Our new video follows one of those questions on a real cluster, and where it led: Are we using our flash tier efficiently? How much data could we shift to HDD? If you want to see more, join us on October 7th, where we go further into the performance work in Quobyte 5 and run the MCP server interface live. Inside Quobyte 5: Turn Your Hardware Into Lightning-Fast AI Storage October 7th, 10 AM PT Register: buff.ly/DDyQ7wk #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage #Quobyte

"What percentage of my data is cold, and how much of it is sitting on flash?" Most storage teams can't answer that without a specialist and half a day of work. In this demo, we asked our file system directly, in plain English, and had the answer in seconds. Then we kept going: how much space is tied up in files nobody has opened since the day they were written? Who are our top five capacity consumers this month? Is anyone using flash to store videos? That's the new MCP server interface in Quobyte 5. It connects an LLM to the Quobyte File Query Engine, so anyone can ask what's on their storage and get answers straight from live metadata. No query syntax, no separate index, and your existing permissions still apply. Watch the demo. If you could ask your file system one question, what would it be? #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage #Quobyte

If you run AI or HPC infrastructure, you know the pressure to get more from the hardware you already own. New hardware takes months to arrive, more servers mean more power, and neither solves the storage bottleneck that leaves your GPUs waiting on data. Quobyte 5 is here, and it is built for exactly that. The x86 and ARM servers you already own can now deliver substantially higher I/O and metadata performance while consuming less power. And with Quobyte GPU Converged Storage, surplus CPU and NVMe in your GPU nodes can also become high-performance, resilient storage. More of your return comes from the hardware already in your racks. Quobyte 5 also introduces a generally available MCP server interface connecting LLMs to the Quobyte File Query Engine. Ask questions in plain language, get answers from live metadata, and decide what to move, what to reclaim, and what to buy next. For example, you can ask what percentage of your data is cold, what is sitting on flash that shouldn't be, and who is consuming your capacity. The answers were always in your file system. Quobyte 5 makes them easier to reach, understand, and act on. Read the full announcement: buff.ly/3lXC0cG #AIInfrastructure #HPC #SoftwareDefinedStorage #AIStorage

Most storage vendors want you to treat their system as a black box. Trust the appliance, file a ticket when it breaks, and wait for a support engineer to explain your own infrastructure back to you. We built Quobyte the opposite way. Administrators can look under the hood: human-readable RPC traces, understandable on-disk formats, and deterministic policy-based data placement instead of hashing algorithms that decide where your data lands and then force a rebalance every time you add a node. That difference shows up the day you expand a live cluster. With explicit placement, you know exactly what happens next. New drives start taking new files within seconds, and nothing else moves. No rebalance storm competing with your training job for the same hardware. At scale, predicting what the system will do separates teams that operate their storage from teams that react to it. Our technology whitepaper covers it all: the fault tolerance model, the client architecture, the policy engine, and the security design. Worth reading if you're making storage decisions at scale: buff.ly/A37l7lk #ArchitectedForAI #Quobyte #AIInfrastructure

Training AI on 500 million medical images sounds impossible, until you fix the storage bottleneck. Siemens Healthineers runs 1,600 deep learning experiments a day on their "Sherlock" supercomputer, but the real challenge was never the GPUs. It was keeping them fed. Copying multi-terabyte datasets to scratch space before every training job just wasn't practical at that scale anymore. So they moved to Quobyte, a parallel file system that lets training jobs read data directly at high speed, no staging required. Automatic tiering keeps hot data on NVMe and cold data on cheaper HDDs, and 3-way replication lets the team survive the loss of two entire storage servers with zero downtime, all running on the commodity hardware they already owned. Worth a read if your AI infrastructure is starting to feel the same strain: buff.ly/yW1wiQm #AIInfrastructure #Quobyte #ArchitectedForAI

Your GPUs may not be waiting on the model. They may be waiting on storage. In multi-turn, agentic LLM systems, every resumed conversation needs its KV cache pulled back from storage before the model can generate the next token. At scale, that reload path can become a major bottleneck, leaving GPUs idle while context moves back into memory. Learn more about why KV cache offloading matters for agentic AI, and how Quobyte GPU Converged storage can make context reload faster, resilient, and available to any GPU node. Read here: buff.ly/eyQKncy #Quobyte #AIInfrastructure #KVCacheOffloading #AgenticAI #LLMs #ArchitectedForAI

Most storage architectures have a ceiling, and you usually discover it just as your AI program starts to scale. Zoox grew from 2 PB to more than 70 PB on the same Quobyte AI storage platform, without rearchitecting the environment or going through a forklift upgrade. Quobyte keeps thousands of GPUs fed with video, lidar, radar, and simulation data, while hybrid storage keeps costs under control. Storage should never be the reason an AI roadmap stalls, and for Zoox, it isn’t. Read the story here: buff.ly/FbaNtzm #AIInfrastructure #Quobyte #ArchitectedForAI

Delivering storage as a service means every tenant should feel like the infrastructure is theirs alone, without deploying a separate storage system for each one. Quobyte's multi-tenancy extends across the entire platform: the file system, S3, the API, the CLI, and the web console. Each tenant sees only their own environment, with their own volumes, users, groups, quotas, and LDAP or Active Directory integration. Shared infrastructure by default; dedicate drives and servers when performance isolation or compliance demands it, without creating another storage silo. #Multitenancy #SaaS #AIInfrastructure #Quobyte

The conversation around AI infrastructure is shifting from training and fine-tuning to inference, particularly multi-turn agentic AI. Every interaction may require a large volume of context, including KV cache and session data, to be loaded quickly. When storage cannot move that data fast enough, GPUs wait. In his latest @BlocksandFiles, @Chris_Mellor explores how Quobyte GPU Converged Storage accelerates this process by using the high-performance network and local resources already available within GPU clusters to deliver context directly into host memory. Read the full story here: buff.ly/tvmkRta #AIInfrastructure #GPUUtilization #AgenticAI

What a few days at ISC. The booth stayed busy from start to finish, the conversations were sharp, and Jimmy the skeleton drew a crowd of his own. Thank you to everyone who came by E01, ran a demo, and said hello. We had a blast, and we are already looking forward to the next one. #ISC26 #HPC #AI #Storage 

Last day of ISC, and the booth comes down tonight. If your GPUs have spent this year waiting on storage, this is your chance to see the fix in person. We are showcasing Storage Architected for AI™ at Booth E01 until the floor closes. #ISC26 #HPC #AI #Storage

The line for a photo with Jimmy has not let up all day, and neither have the conversations behind it. Plenty of people here know the feeling of paying for GPUs that spend half their time waiting on storage. Come by Booth E01 and we will show you how good it can be instead. #ISC26 #HPC #AI #Storage

The fastest GPUs are only as fast as the data you can get to them. Starve them and you pay top dollar for silicon that sits idle. Quobyte keeps the pipeline full so your GPUs run flat out. See it at E01. Jimmy is hard to miss, the demo is worth the stop. #ISC26 #HPC #AI #Storage

The fun has begun at Booth E01, and Jimmy the skeleton already has a line forming for photos. He is well on his way to most photographed attendee at ISC. Come get yours, grab some swag, and say hi to the team. #ISC26 #HPC #AI #Storage